Software engineer building agent infrastructure: orchestration layers, safe local execution, typed memory, routing policy, evaluation, and the operator surfaces that make autonomous runs accountable.
I care about the unglamorous parts of agent systems: the queue, trace, sandbox, review gate, memory record, cost ledger, and handoff. If those pieces are weak, the model looks impressive right up until it repeats the same expensive mistake.
| Surface | Maturity | What to inspect |
|---|---|---|
| JustAi | Prototype | Intent, planning, review, and checkpoint control plane. The public execution path is currently unwired and fails closed. |
| JustAi demo | Simulation | Deterministic, fixture-backed mission-control walkthrough. It has no backend or live agents. |
| safe-mini | Alpha substrate | Policy-guarded local bash-action loop. Built-in executors run in the host process; isolation requires an injected executor boundary. |
| route-mini | Reference | Deterministic multi-provider routing policy with fallback, budget, latency targets, and decision records. |
| memory-mini | Reference | Namespaced memory semantics: upsert, soft delete, cleanup, retrieval, and optional embeddings. |
| lab-mini | Reference | Repeatable data-science lab loop: load, profile, analyze, claim, report. |
These labels describe the public artifacts as they work today, not the larger architecture they are intended to become.
Agents get useful when the system around them is engineered like production infrastructure.
- Sandbox the boundary, not the capability. Give the agent room to solve the task inside a scoped worktree, scrubbed environment, guarded path, and recorded trajectory.
- Trajectories beat vibes. Every action should leave a replayable trace that can teach the next run.
- Routing is policy. Stronger models are an escalation decision, not a default reflex.
- Memory has lifecycle. Durable context needs namespacing, upsert, retention, and cleanup rather than chat-history luck.
- Evaluation should change behavior. A score that does not route, block, or teach the next run is mostly decoration.
- Portfolio and writing: delegateandorchestrate.com
- Interactive simulation: justai-demo.vercel.app
- Control-plane prototype: JustAi
- Substrate and reference repos: safe-mini, route-mini, memory-mini, lab-mini
I keep forks and notes around projects that shape the work:
- mini-swe-agent and SWE-agent for minimal bash-action agents and issue-solving loops.
- MCP Python SDK and MCP TypeScript SDK for tool protocol boundaries.
- smolagents and Vercel AI SDK for code-agent and TypeScript agent interface patterns.
The goal is not to collect logos; it is to keep the public graph close to the ideas I am building against.
- US Navy veteran, aircraft maintenance. The useful lesson was not ceremony; it was operational seriousness.
- Full-stack engineer across Python, TypeScript, React, Node, SQL, local Linux/WSL, and browser-visible product surfaces.
- Building Delegate & Orchestrate around practical agent systems that can be inspected, tested, and improved.